ACCURACY OF HOSPITAL DISCHARGE ABSTRACTS FOR IDENTIFYING STROKE

Background and Purpose Much of the available data on stroke occurrence, service use, and cost of care originated with hospital discharge abstracts. This article uses the unique resources of the Rochester Epidemiology Project to estimate the sensitivity and positive predictive value of hospital discharge abstracts for incident stroke. Methods The Rochester Stroke Registry was used to identify all confirmed first strokes (hospitalized and nonhospitalized) among Rochester residents for 1970, 1980, 1984, and 1989 (n=364). The sensitivity of discharge abstracts was estimated by following these individuals for 12 months after stroke to determine the proportion assigned a discharge diagnosis of cerebrovascular disease (International Classification of Diseases [ICD] codes 430 through 438.9). The positive predictive value of discharge abstracts was assessed by identifying all hospitalizations of Rochester residents with an ICD code of 430-438.9 in 1970, 1980, and 1989 (n=377). Events were categorized as incident stroke, recurrent stroke, stroke sequelae, or nonstroke after review of the complete community-based medical record by a neurologist. Results Only 86% (n=313) of all first-stroke patients in 1970, 1980, 1984, and 1989 were hospitalized. Of hospitalized patients, only 76% were assigned a principal discharge diagnosis code of 430-438.9. Fatal strokes and those occurring during a hospitalization were less likely to be identified. Among all hospitalizations of Rochester residents in 1970, 1980, and 1989, there were 377 with a principal diagnosis code of 430-438.9. Less than half (n=177) were determined by the neurologist to be incident stroke; only 60% (n=225) were either incident or recurrent stroke. Comparison of alternative approaches showed the validity of discharge abstracts was enhanced by increasing the number of diagnoses and excluding codes with poor positive predictive value. Conclusions This study provides previously unavailable estimates of the sensitivity of stroke-coded hospitalizations for a US community. A model for improving the sensitivity and positive predictive value of discharge abstracts is presented.